WebBrain Compass Tiny XS v3
WebBrain Compass Tiny XS v3 is a compact language model optimized for low-latency decision making, structured tool use, and in-browser agentic execution inside the WebBrain runtime.
Fine-tuned from Spark-X2.5-1.7B (~1.7B parameters) (XHToken/Spark-X2.5-1.7B at 14d6e83c13c7add2b62a7c39b2131f4ed1cddcf8), this is the BF16 reference model with merged weights, accompanied by the original LoRA adapter under adapter/.
It unifies three core WebBrain behavioral modes:
- Ask & Clarify — Answer directly or request missing information when execution is underspecified or unnecessary.
- Direct Compact Tool Execution — Select exact browser tools and generate grounded arguments for low-latency accessibility-tree actions.
- Safe Escalation & Abstention — Refuse, pause, or defer to higher execution tiers when actions lack adequate grounding or violate safety boundaries.
Note: Display/repository name is WebBrain Compass Tiny XS v3. This is a metadata-only rename preserving tested weight/code revision
main. For native in-browser WebGPU execution, see the companion ONNX package:webbrain-one/webbrain-compass-tiny-xs-v3-onnx.
Role in WebBrain
User request
│
▼
WebBrain observation & policy layer
│
▼
WebBrain Compass Tiny XS v3 (BF16 Reference)
├─ Clarify or answer directly
├─ Emit grounded browser tool call
└─ Abstain / safely escalate when execution is ungrounded
The outer WebBrain runtime enforces tool-schema validation, evidence and parameter grounding, destination URL verification, browser-state freshness, and sandboxed security policies. Model output is never proof that an external action succeeded.
Intended Capabilities
- Browser Action Selection: Grounded accessibility-tree interactions and structured function calling.
- Reference Decision Making: Fast, high-fidelity local inference for compact browser actions.
- Unified Ask and Compact Modes: Asking for missing information before acting, answering direct questions, or choosing tools.
- Safe Refusal & Escalation: Refusing unsupported actions or escalating when Compact execution lacks evidence, avoiding invented URLs or fabricated success states.
Quickstart & Loading
You can load and run WebBrain Compass Tiny XS v3 directly with Hugging Face transformers:
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
repo = "webbrain-one/webbrain-compass-tiny-xs-v3"
revision = "main" # Tested weight/code revision
tokenizer = AutoTokenizer.from_pretrained(repo, revision=revision)
model = AutoModelForCausalLM.from_pretrained(
repo,
revision=revision,
trust_remote_code=True,
dtype=torch.bfloat16,
device_map="auto"
)
# Example structured tool use
messages = [
{"role": "user", "content": "Click the sign-in button on the page"}
]
inputs = tokenizer.apply_chat_template(
messages,
tools=tools,
enable_thinking=False,
add_generation_prompt=True,
return_tensors="pt",
return_dict=True
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False)
Tool Calling Format
Use native structured tool_calls with dictionary arguments; do not reuse MiniCPM's template or tool serialization. Spark emits <tool_call>name with <arg_key>/<arg_value> fields. For long contexts in production, use a Spark-compatible serving runtime.
Technical Specifications
Fixed BF16 Routing Results (2026-09-26)
| Metric | BF16 Routing Score |
|---|---|
| First-turn structured calls | 97/100 |
| Strict exact action | 9/89 |
| Loose tool-family match | 39/89 |
Evaluated across 100 first turns and 89 scored fixed-history scenarios (11 predetermined skips). Retains five generation timeouts and five discouraged actions in the denominators without cherry-picking.
License
Noncommercial research only. This model and its fine-tuned weights are for noncommercial research purposes only. The upstream Spark model/code license is included in LICENSE; it does not change this release's usage restriction.